On October 6, 2025, at OpenAI DevDays, CEO Sam Altman unveiled AgentKit, a game-changing platform that transforms the traditionally complex, fragmented process of building intelligent agents into a seamless, visual development experience. But what makes this announcement even more compelling for developers is how quickly the ecosystem is expanding with powerful integrations.
The Road So Far…
Every agent project typically starts with some considerable chunk of time spent on boilerplate setup. Developers have to manually configure multiple SDKs, manage dependencies, and wrangle sometimes incompatible API versions. A simple “Hello World” agent often requires 100+ lines of configuration code before writing a single line of business logic.
Moreover, getting agents into production requires a tinge of DevOps expertise most developers don’t want to touch, which includes build pipeline, deployment configurations, scaling policies, monitoring setup, and error handling. Not to mention that not everyone is a fan of Python or TypeScript, the languages used at the backend.
As agents evolve, so does their complexity. Adding new capabilities means refactoring core architecture, updating multiple integration points, and praying nothing breaks in production. A simple feature request can cascade into weeks of technical debt repayment.
So far, building effective agents has required deep expertise across multiple domains, including ML model integration, API orchestration, conversational design, and infrastructure management. Most development teams had gaps in at least one critical area.
AgentKit’s Approach
AgentKit addresses these implementation challenges through several interesting architectural decisions:
- Visual Development Interface
- Drag-and-drop component assembly
- Real-time workflow visualization
- Configuration through UI forms rather than code
- Use of the OpenAI Agents SDK as a backend
What OpenAI has done is build on top of the earlier 2025 releases of Responses API and Agents SDK, and package the logic into a visual builder that looks to be quite the player in the already crowded field. Those two already include the support for invoking external tools via MCP, so all the ingredients are here.
In contrast to what you can get on other AI vendor-powered platforms on the market, like Copilot Studio or Google Agent Studio, this one seems like a more rounded package because it also includes first-party evals and tracing of the Agent conversations.
The Reality Check
AgentKit definitely reduces the time spent on reading documentation and debugging configuration syntax. It also makes agent logic more accessible to non-backend developers.
However, visual tools can become unwieldy for complex logic, and AgentKit may not be able to support advanced customization scenarios that require direct code access. The ability to inspect and export the code generated by the visual builder is a great feature, especially since it allows you to toggle between Python and TypeScript variants. Nevertheless, for complex customizations, most developers will likely treat AgentKit as a starting point, using it to generate an initial code structure before moving development outside the platform for more granular control.
The same goes for third-party integrations. While it eliminates integration boilerplate and reduces time-to-integration for supported services, it’s still limited to a bunch of supported integrations. Custom or niche services may still require manual SDK work.
The major missing link here is also the bring-your-own-model feature, which means you are doomed to using OpenAI’s models, which, while definitely great, are not the best-fitting solution for every use case. This will likely be addressed in the long term, as the underlying framework already includes the capability to connect to any OpenAI-like endpoint. For now, visually, the option just isn’t there.
One thing’s for sure. AgentKit’s approach represents a notable shift toward accessible agent development, though it’s still early to assess its long-term impact on the development ecosystem.
Market Implications and Future Outlook
Major players on the market that are betting exclusively on the conversational AI (like kore.ai) and/or products that are just a thin wrapper around OpenAI’s models are in danger of losing customers in the long run.
The platform’s success will largely depend on how well it balances ease of use with the flexibility developers need for complex applications, and whether OpenAI can maintain a sustainable cost structure while expanding integration capabilities.
Time will tell how the developer community adopts these tools and what direction the platform evolves in, but the initial architectural decisions suggest a promising foundation for reducing the traditional barriers to agent development.
MCP Servers to the Rescue
While AgentKit provides solid built-in integrations, real-world applications often demand specialized capabilities that go beyond standard offerings. Model Context Protocol (MCP) servers effectively address this need, and their integration with AgentKit works well.
The MCP integration process is more straightforward than traditional SDK integrations. Once an MCP server is configured and connected, AgentKit treats its capabilities as native tools, creating a seamless development experience. The standardized protocol means less time spent on authentication patterns and API quirks that typically complicate custom integrations.
Infobip MCP Integration
Good news for those who want to add communication superpowers to their agents. Our engineering team has successfully integrated our own Infobip MCP Servers with OpenAI’s Agent Builder.
Supported Communication Capabilities
- Send messages, using various channels, such as SMS, WhatsApp, Viber, RCS
- Set up and manage 2FA
- Access your customer data platform
- Manage Infobip user accounts
- Explore Infobip Documentation
Infobip MCP Servers provide several key benefits for developers building communication-enabled agents. The servers eliminate the need to integrate and manage separate SDK implementations for each communication service. Real-time delivery tracking provides status updates for sent messages, allowing agents to respond appropriately to delivery failures or delays. Template management handles dynamic content generation and compliance requirements automatically, which is particularly valuable for regulated channels like WhatsApp Business. Built-in error handling includes automatic retry logic and fallback mechanisms, ensuring message delivery reliability without requiring custom error management code.
Conclusions
AgentKit represents a significant shift toward accessible agent development, with clear benefits for rapid prototyping and standard use cases. The addition of MCP servers extends these capabilities into practical business applications.
For developers interested in communication-enabled agents, the combination of AgentKit’s development experience and Infobip’s MCP integration provides a practical starting point with production-ready messaging capabilities.
